See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/274249921

Automatic labeling of phonesthemic senses.

Conference Paper · January 2013

CITATIONS READS
13 282

3 authors:

Ekaterina Sangati Raquel Fernández

Okinawa Institute of Science and Technology University of Amsterdam
13 PUBLICATIONS   114 CITATIONS    80 PUBLICATIONS   777 CITATIONS   

SEE PROFILE SEE PROFILE

Federico Sangati
University of Naples "L'Orientale"
40 PUBLICATIONS   211 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Emojiworldbot View project

PARSEME-IT - Syntactic Parsing and Multiword Expressions in Italian View project

All content following this page was uploaded by Ekaterina Sangati on 23 June 2016.

The user has requested enhancement of the downloaded file.

 Automatic Labeling of Phonesthemic Senses
Ekaterina Abramova (e.abramova@ftr.ru.nl)
Department of Philosophy, Radboud University Nijmegen

Raquel Fernández (raquel.fernandez@uva.nl)

Institute for Logic, Language & Computation, University of Amsterdam

Federico Sangati (federico.sangati@gmail.com)

Institute for Logic, Language & Computation, University of Amsterdam

Abstract
This study attempts to advance corpus-based exploration of
sound iconicity, i.e. the existence of a non-arbitrary relation-
ship between forms and meanings in language. We examine a
number of phonesthemes, phonetic groupings proposed to be
meaningful in the literature, with the aim of developing ways
to validate their existence and their semantic content. Our
first experiment is a replication of Otis and Sagi (2008), who
showed that sets of words containing phonesthemes are more
semantically related to each other than sets of random words.
We augment their results using the British National Corpus and
the Semantic Vectors package for building a distributional se-
mantic model. Our second experiment shows how the semantic
content of at least some phonesthemes can be identified auto-
matically using WordNet, thereby further reducing the room
for intuitive judgments in this controversial field.
Keywords: Iconicity; Phonesthemes; Corpus analysis; Distri-
butional Semantics; WordNet.
Introduction
The claim that the relationship between forms and meanings
in language is not always arbitrary is controversial. How-
ever, evidence for non-arbitrary relationships comes at mul-
tiple levels of language, from phonology to syntax (Perniss,
Thompson, & Vigliocco, 2010). Here we focus on the pho-
netic level and investigate the association of particular sounds
with aspects of word meaning. Such sound iconicity has been
described in a variety of non-Indo-European languages (see
the studies in Hinton, Nichols and Ohala Hinton, Nichols,
& Ohala, 2006b) and its existence in English suggested by
a number of authors (Firth, 1930; Marchand, 1969), and ex-
ploited in commercial settings (Shrum & Lowrey, 2007).
Phonesthemes (a technical term for meaningful sound pat-
terns) are sub-morphemic units that play a role of morphemes
but have been traditionally distinguished from them by being
non-compositional (but see Rhodes Rhodes, 2006 for an op-
posite view). The most oft-cited example is the English phon-
estheme gl which occurs in a large number of words related to
light or vision (glitter, glisten, glow, gleam, glare, glint, etc.).
Once the phonestheme is taken out, the remainder of the word
is not a morpheme (-itter, -isten, -ow etc.) and one does not
attach gl to other words to make them light-related. Still, the
extent and the nature of this phenomenon is not clear.
Traditionally, the evidence for the existence of phones-
themes and their proposed meaning consisted in listing a
number of words that share a given sound and attempting
to find the semantic core that unites them. Popular expla-
nations for the phenomenon would rest on the intuited as-
sociation between sound production and meaning. For ex-
ample, Reid (1967) states that “The explosive nature of the
letter b is intensified when it is combined with l before the
breath is released. Consequently words beginning with bl are
found generally to indicate a ’bursting-out’ or the resultant
swelling or expansion” (p. 10). More recent accounts view
them rather as a matter of statistical clustering. According to
such “snowballing effect” theory, a group of phonemes in re-
lated words (for example, by common etymology) becomes
over time associated with the meaning of these words and
given the right conditions starts to attract other words with
the same phoneme into a cluster, through semantic change or
influencing the creation of new words (Blust, 2003; Hinton,
Nichols, & Ohala, 2006a).1
Dissociating these competing explanations would require
a combination of historical and cross-linguistic research but,
arguably, there is a wealth of more basic questions that need
to be addressed first. The nature of iconicity is such that it is
easy to see the connection between form and meaning once
we are aware of both elements but such intuition is not al-
ways a reliable guide for discovering the connections. Just as
it is difficult to interpret an iconic sign from American Sign
Language when its meaning is unknown (Bellugi & Klima,
1976), we might miss the connection that is in fact present.
On the other hand, we might over-estimate the connection
by listing only the light-related gl-words and forgetting the
amount of gl-words that have nothing to do with light (glide,
glucose, globe, glove, etc.). In other words, if we want to val-
idate the existence of phonesthemes or explain their origin,
we need to apply more falsifiable and unbiased methods in all
stages of investigation: identifying them in a given language,
quantifying their scope and establishing their meaning.
So far, the reality of phonesthemes has been demonstrated
in behavioral experiments (Bergen, 2004; Hutchins, 1998)
and corpus studies (Drellishak, 2006; Otis & Sagi, 2008). Our
aim is to contribute to the second current of this research. We
believe that this is a valuable way of objectively addressing
large-scale linguistic phenomena that can refine our under-
standing of sound iconicity and lead to further testable hy-
1 This is not to say that there are no universal sound features un-
derlying certain cases of sound iconicity, such as words for small
and large objects usually associated with high and low acoustic fre-
quency respectively (Ohala, 1994).
potheses with respect to its cognitive underpinnings.
Otis and Sagi (2008) conducted the first corpus-based anal-
ysis of phonesthemes. They examined 47 groups of words
containing phonesthemes using Project Gutenberg texts and
a method for calculating word similarity based on Latent
Semantic Analysis (LSA), in particular its Infomap2 variety
(Schütze, 1997). The analysis performed by Otis and Sagi
showed that semantic relatedness of clusters of words that
share a phonestheme is higher than that of clusters composed
of randomly chosen words. This method, therefore, can be
used to examine the validity of conjectured phonesthemes.
However, as the authors admit, it “does not identify what
specific semantic content is carried by the identified phon-
estheme” (p. 68). Our first aim is replicating the study of Otis
and Sagi using (1) a more recent and balanced corpus – the
British National Corpus (BNC), and (2) a newer and more
versatile and efficient tool for calculating semantic related-
ness, Semantic Vectors3 (Widdows & Cohen, 2010).
Our second aim is attempting to develop a method for au-
tomatically identifying the semantic content associated with
a particular phonestheme—a task that, to our knowledge, has
not previously been addressed in the literature. Otis and Sagi
(2008) suggest that methods designed to identify the topic of
a given text could be used to that end. We think, however, that
a more straightforward method lies in analogy with the task
of unsupervised ontology acquisition: placing a word within a
hierarchy of concepts based on its semantic relationship with
the rest of the words in the hierarchy: for example, pear be-
ing placed close to apple and banana under fruit. In the case
of phonesthemes, it is conceivable that a group of gl words
would be assigned a vision-related higher class. Whether this
can be done automatically and applied to a variety of phones-
themes is one of the questions we pose in this study.
In sum, our hypotheses are the following:
Hypothesis 1: Words that share a phonestheme are on av-
erage more semantically related than random words.
Hypothesis 2: The core semantic import conjectured in
the literature for a phonestheme can be derived automatically
from a set of phonestheme-bearing words.
Experiment 1: Semantic Relatedness
Methods
To explore our Hypothesis 1, we used the British National
Corpus (Burnard, 2007), a 100 million word collection of
written and spoken English language compiled from a wide
variety of sources and genres. We pre-processed the entire
corpus using the Natural Language Toolkit (NLTK) (Bird,
Klein, & Loper, 2009). In particular, we used NLTK to ex-
tract the content words in the corpus (nouns, adjectives, and
lexical verbs) and to lemmatize them, i.e. to reduce a family
of inflected words (such as walk, walks, walked, walking) to a
single word type or lemma (e.g. walk). This resulted in a sub-
corpus of about 43 million words, which we used as input to
2 Freely available at http://infomap-nlp.sourceforge.net.
3 Freely available at http://code.google.com/p/semanticvectors/.
construct a distributional semantic model with the Semantic
Vectors package (Widdows & Cohen, 2010).
Semantic Vectors allows us to use a corpus to build a
high-dimensional vector space where words are represented
as vectors that record their frequency of co-occurrence with
other words or other documents in the corpus. We can then
use well-defined methods to measure how similar the mean-
ings of two words are, such as computing the cosine of the
angle formed by their corresponding vectors. As Otis and
Sagi (2008) indicated in their pioneering corpus study, this
methodology can be of great value to investigate the claims
behind phonesthemes in an objective, data-driven way, since
we can use the distributional model to test whether words
sharing a hypothesized phonestheme exhibit higher semantic
similarity than random words.
Like Otis and Sagi, we built a term-term model where each
word vector records the co-occurrence of that word with other
words in the context (rather than recording occurrence in par-
ticular documents like LSA), but unlike them, who used the
traditional singular value decomposition method for reducing
the dimensions in the matrix, we used Random Projection, a
more computationally efficient algorithm.4 We experimented
with the settings of two parameters in the Semantic Vectors
package: the minimum frequency of the word types consid-
ered for building the model (as we may not be able to con-
struct reliable distributional semantic representations for low
frequency words) and the window size, i.e. the context win-
dow of n words to left and right of the target word where the
model looks for co-occurrences of other words. McDonald
and Ramscar (2001) claim that “the best fit to psychological
data is typically achieved with word vectors constructed us-
ing context window sizes between ±2 and ±10 words.” Otis
and Sagi used n = 15, which is the default setting in Infomap.
We focused on the 22 prefix phonesthemes conjectured by
Hutchins (1998). Our statistical analysis followed the pro-
cedure proposed by Otis and Sagi (2008). For each phones-
theme, we first extracted all the vectors of the phonestheme-
bearing word types in our distributional semantic model.5 We
shall refer to the resulting set of words (and vectors) as a
phonestheme cluster. We then performed two Monte Carlo
analyses. In the first analysis, we computed the average se-
mantic similarity of each phonestheme cluster by forming
1000 random pairs and averaging the semantic distance ob-
tained. In addition, we did the same for similarly-sized clus-
ters of random words and performed an independent sam-
ples t-test for the resulting two groups of values. In the sec-
ond analysis we took 50 random pairs within each phones-
theme cluster and a corresponding group of pairs of random
words and run 100 independent sample t-tests noting whether
the mean of phonestheme cluster distances was significantly
higher than the distances obtained for pairs of random words
(with α = 0.05). Based on the binomial distribution, we
4 See Sahlgren (2005) and Widdows and Cohen (2010) for a com-
parison of these methods.
5 Since we are dealing with a written corpus, this is done on the
basis of an orthographic match with the phonetic grouping.
judged the number of significant t-tests as higher than 15 to average semantic relatedness of phonestheme clusters (M =
lend statistical support to our Hypothesis 1. We performed 0.446, SD = 0.044) was highly correlated with the number of
the procedure 5 times and took the mean to be the final result. significant t-tests (r = 0.95, p < 0.0001) and was furthermore
We used the results obtained with the gl phonestheme clus- significantly higher than the average semantic relatedness of
ter (which obtained the highest statistical support in the Otis random words clusters (M = 0.397, SD = 0.018), as shown
and Sagi study) to optimize the minimum frequency and by an independent samples t-test (t(42) = 4.83, p < 0.0001).
the window size parameters of our distributional semantic In line with the findings of Otis and Sagi (2008), we were
model. The model produced the most qualitatively sensible thus able to obtain support for Hypothesis 1 for 16 conjec-
and most statistically stable results when setting the minimum tured phonestheme prefixes. Using the BNC – a more gen-
frequency to 100 and the window size to 10. This resulted in eral, balanced, and modern corpus of English than Project
a model containing a set of 22292 vectors. This vector space Gutenberg – our study yielded higher support for the hypothe-
was used in all subsequent parts of our study. sis than Otis and Sagi’s previous study, which had found only
12 phonestheme prefixes as reaching statistical significance.
Results
The results obtained with our parameter optimizing test on Experiment 2: Phonestheme Cluster Labeling
the gl phonestheme showed that the semantic relatedness of
Methods
the words in the gl cluster was significantly higher than that
of clusters of random words, as measured by our t-test pro- After establishing significant differences between the seman-
cedure. On average, 26.4 t-tests produced a significant result tic relatedness scores for phonestheme word clusters and
(recall that the threshold of significance of the binomial test clusters of random words, we turned to our second experi-
was for at least 15 out 100 t-tests to turn out as significant). ment, whose aim is to investigate possible methods to auto-
We used the same vector space (with the parameters fixed) matically detect the core semantic content carried by a phon-
to analyze the remaining 21 prefix phonesthemes. The results estheme. To our knowledge, this is the first attempt to ad-
obtained are reported in Table 1. For each phonestheme, the dress this issue by objective means. To test our Hypothesis
table shows the number of word types in the phonestheme 2, we selected a number of prefix phonesthemes based on the
cluster (# Tokens), the average degree of semantic related- amount of statistical support obtained in our first experiment
ness amongst those words (Sim) calculated according to our and on how unambiguous and generally agreed upon were
first Monte Carlo analysis, the number of significant t-tests the sense definitions proposed in the literature. We selected
(# Sig) calculated according to our second Monte Carlo anal- 10 phonesthemes with high semantic relatedness scores and
ysis, and the mean effect size (Effect) of these t-tests. As compiled a list of definitions based on the descriptions given
can be seen, the model did not only confirm the semantic by Hutchins (1998), Marchand (1969) and Reid (1967). The
similarity of the words in the gl phonestheme (for which it resulting list of phonesthemes together with their conjectured
had been optimized), but produced significant results for 16 semantic import is presented in Table 2.
different prefix phonesthemes out of the 22 considered. The
Table 2: Phonesthemic senses
Prefix Definition Example
Table 1: Phonestheme semantic relatedness results
bl- swelling, explosion, extension, broadness bloating
Prefix #Tokens Sim #Sig Effect gl- light, vision, look, brightness, shine glitter
bl- 105 0.4607 56.8 0.2845 gr- threatening noise, anger, grip growl
cl- 156 0.4295 29.4 0.2570 scr- unpleasant sound, irregular movement screech
cr- 197 0.3921 7.40 0.2327 sn- nose, mouth, smell, snobbish person sneeze
dr- 99 0.4504 63.6 0.2849 spl- divergence, spread, splash splash
fl- 137 0.4340 34.2 0.2536 squ- discordant sound, softness, compression squeeze
gr- 197 0.4050 25.2 0.2617 str- linear, forceful action, effort strike
sc-/sk- 167 0.4031 10.2 0.2443 sw- rhythmical movement swing
scr- 32 0.5174 68.4 0.3093 wr- irregular motion, twist wring
sl- 83 0.4275 42.4 0.2734
sm- 42 0.4803 51.4 0.2817
sn- 40 0.4650 52.2 0.2909 In order to automatically assign a semantic class label to a
sp- 161 0.4127 14.4 0.2392 phonestheme cluster, we used WordNet (Fellbaum, 2005), a
spl- 11 0.5224 59.0 0.2723 cognitively motivated ontology of words and concepts linked
spr- 24 0.3950 4.80 0.2373
squ- 24 0.4916 67.0 0.3205
by different semantic relations commonly used in compu-
st- 298 0.4307 25.0 0.2465 tational linguistics. The main semantic relation connecting
str- 89 0.4525 61.8 0.2899 words that express different concepts in WordNet is the su-
sw- 67 0.5138 91.2 0.3396 per/subordinate relation (also called hypernymy/hyponymy),
tr- 249 0.3912 5.80 0.2318 which establishes a hierarchy of concepts from more general
tw- 22 0.4304 17.4 0.2408
wr- 38 0.5155 90.6 0.3915 concepts like animal to increasingly specific ones like mam-
mal or whale. Since hypernymy is a transitive relation, for
each word we can construct its hypernymy chain: the set of
all its superordinate concepts or hypernyms connecting the
word in question to the root node in the hierarchy (entity in
the case of WordNet), ordered by their level of specificity.
WordNet is made up of independent hierarchies for dif-
ferent parts of speech: nouns, verbs, and adjectives. Given
this (which prevents the possibility of assigning a cross-
categorical semantic label) and the fact that the hierarchies
for verbs and adjectives are far less complete than the noun
hierarchy, we focused on the common nouns6 within each
phonesthemic cluster amongst those listed in Table 2. This
resulted in eliminating 37% of words over all clusters.
For each common noun w in a phonesthemic cluster, we
computed a set H(w) containing all superordinate concepts
in the hypernym chain of w, and derived a set H of poten-
tial class labels for that cluster by taking the union of all sets
H(w) for each noun w in the cluster. We then considered sev-
eral methods for selecting the most optimal semantic class
labels from H . Our methods were inspired by the approach
to unsupervised ontology acquisition proposed by Widdows
(2003) according to which “the most appropriate class-label
for the set [of words] S is the hypernym h ∈ H which sub-
sumes as many as possible of the members of S as closely as
possible in the hierarchy” (p. 278). Widdows offers a general
scheme for defining an affinity score function α(w, h) between
a word w and a candidate label h, which generates a ranking
of all the potential class labels for a cluster of words:
(
f (dist(w, h)) if h ∈ H(w)
α(w, h) =
−g(w, h) if h ∈
/ H(w)
where dist(w,h) is a distance measure between a given word
and a hypernym, f is a reward function that gives points to h if
it subsumes w and the more points the closer this relationship,
while g is a penalty function that subtracts points if h does not
subsume w. The best class-label is the hypernym h ∈ H that
has the highest affinity score summed over all the elements in
the cluster.
Following Widdows, we chose as our distance measure the
number of intervening levels in the WordNet hierarchy and
set the rewarding function to f = 1/dist(w, h)2 . As for the
penalty function g, we tested constant values of 0.25, 0.1 and
0.01. This particular variant of the scoring function thus mag-
nifies the credit given to classes that are close to the words
they subsume while giving a very small penalty to potential
labels that miss out words in the cluster. The expected result
is thus a ranking of class labels with a strong preference for
specificity. This seems congruent with the nature of phon-
estheme clusters, which may contain a relatively large num-
ber of words that due to, for example, etymological factors
are unlikely to be all related to the phonesthemic meaning.
In fact, it is acknowledged in the literature that the sound-
meaning associations are likely to be probabilistic (Hutchins,
6 We discarded proper nouns, which in WordNet are always ter-
minal leaf nodes representing concrete instances rather than types.
1998) and that phonesthemic meaning can fall into related
but separate groups. For example, gr is taken to be related to
both angry noises (growl, grunt) and grabbing actions (grab,
grasp). Given this, we also considered an approach whereby
we first run a Gaussian Expectation-Maximization clustering
algorithm on each phonestheme cluster to obtain more refined
subsets of words and then run our scoring function algorithm
on each of the resulting sub-clusters.
Finally, to counterbalance the preference for high speci-
ficity but potentially low coverage of the words in the phon-
esthemic clusters, we experimented with a different labeling
algorithm that fixed a minimum coverage threshold. The al-
gorithm examines all hypernyms h ∈ H , selects those that
subsume a minimum percentage θ of words in the cluster and
then ranks them according to their specificity (the number of
intervening levels to the root node entity). We tested the per-
centage values θ = 10 and θ = 20 and run the algorithm both
on complete cluster phonesthemes and on the unsupervisedly
derived sub-clusters.7
Results
Our results show that successful labeling of phonesthemic
clusters can be performed but success depends on a number of
factors. First, it is necessary to clarify what we mean by suc-
cessful labeling. A labeling outcome of phonesthemic senses
was deemed successful when the top 10 labels fulfilled the
following criteria:
1. the topmost label is not the WordNet root node (entity);
2. the top 5 labels do not all have specificity score m ≤ 2;
3. at least 50% of the top 10 labels carry meaning predicted
for a given phonestheme;
4. the top 10 labels together subsume at least 50% of the
words in the cluster.
These heuristics mean that if it is possible to establish the
semantic core of a phonesthemic cluster using WordNet hy-
pernym trees, the top labels will be both specific and in the
direction predicted by the literature. It is always possible to
subsume all the words in the cluster, independently of their
semantic relationship, under the root, just due to the Word-
Net structure. Such a label, however, would not be very in-
formative. By the same reasoning, we excluded the next two
levels of the hierarchy which contain concepts such as physi-
cal entity, abstraction, matter or relation. On the other hand,
specificity needs to be balanced out by coverage, i.e. it is pos-
sible to have very specific labels as top results but covering
only a small portion of words in the cluster. Finally, the la-
bels need to at least intuitively relate to the domain specified
in the literature for a given phonestheme.
Given these criteria, we obtained clear positive results for
one phonestheme (gl) out of 10 examined; moderately suc-
cessful results for two phonesthemes (sn and str); and neg-
ative results for the remaining 7 phonesthemes. We present
7 Assigning a higher value to g would also increase coverage.
However, for consistency g would have to be dependent on the size
of the cluster. We instead choose a simple approach here which re-
sorts to a percentage.
 presence of two clusters for 4 out of 10 phonesthemes we
Table 3: Top 5 WordNet labels for gl-, sn-, and str-
considered (bl, gl, gr and str).8 Again, however, the most
Prefix Label Score Spec Cov interesting result was obtained for the gl phonestheme. Ac-
gl- brightness 4.82 6 23.7% cording to the labels we obtained for its two sub-clusters,
(N=56) flash 4.67 5 13.2% only one of them was light-related. The second sub-cluster
radiance 3.92 7 13.1% contained words like gluten, glucose and glycoprotein, which
light 2.23 5 26.3%
look 2.16 8 10.5% were placed under labels such as protein, macromolecule and
sn laugh 1.71 5 6.5% organic compound, indicating a clear presence of chemistry-
(N=34) unpleasant person 1.71 8 6.5% related words. No clear patterns were obtained for the other
photograph 1.71 7 6.5% 3 phonesthemes with two sub-clusters. We do not exclude
smell 1.71 8 6.5% the possibility that this result was due to the quality of the
piece 1.71 4 6.5%
noise 1.71 6 6.5% vectors given to our clustering algorithm or to the algorithm
str effort 2.22 8 11.3% itself. For example, the presence of a large proportion of
(N=76) motion 1.14 7 11.3% kinship concepts in one of the gr sub-clusters (grandfather,
labor 0.39 7 11.3% grandchildren etc.) led to it being assigned labels such as
change 0.98 6 20.5% grandparent and ancestor, while the same sub-cluster con-
tained words such as growl and grunt. Therefore, whether
the top labels for the 3 successfully labeled phonesthemes in the sub-clustering step is theoretically sound and if so how it
Table 3, together with the scores calculated by the affinity should be accomplished requires further study.
score function with penalty set to a constant g = 0.01 (Score),
specificity of each label (Spec) and the proportion of words Discussion
in the cluster subsumed (Cov). The results of our first experiment are largely in line with
The gl phonestheme received light- and vision-related la- those of Otis and Sagi (2008), but we also see a number
bels in all labeling algorithms that we tested. As can be seen of differences. On the one hand, we obtain higher support
in Table 3, they are clearly specific, cover a large proportion for phonestheme clusters overall and show statistical signifi-
of words and all carry the predicted meaning. Similar results cance for several phonesthemes previously unsupported. On
are obtained using other two settings of the g function and the other hand, our support for the strongest phonesthemes in
the coverage-based algorithm, although a small percentage of the original study (gl and spr) is weaker. These differences
high-level labels like entity does appear in these lists. can be due to several factors. First, we use a different, more
Our moderately supported phonesthemes sn and str obtain modern and balanced, BNC corpus and our resulting phones-
labels in the predicted direction only with algorithms that theme clusters are larger. Second, we use a different method
reward specificity over coverage. For sn (related to nose, for building our distributional model – both a different algo-
mouth and snobbism according to the literature), the best re- rithm (Random Projection) and a smaller window size.
sult is obtained with the distance measure and penalty func- It is worth noting that our tuning experiments with the gl
tion g = 0.01, while for the str phonestheme, related to force- phonestheme show that the kind of pre-processing that we ap-
ful action – with the coverage-based algorithm of θ = 10. ply to the corpus and the window size parameter do make a
Similar conclusions can be drawn from phonesthemes that difference to the statistical results that can be obtained from
did not lead to clear tendencies in their labels or to specific the model. However, while pre-processing could be viewed
enough labeling. Such lack of success is evident in either as merely a methodological challenge common to all types of
only one label out of the top 10 being relevant to the pre- corpus analyses, there might be a theoretical significance be-
dicted meaning or all of the labels being very general. In the hind the impact of the window size. Sahlgren (2008), for ex-
first case, for example, both gr and scr words are subsumed ample, suggests that a small window size is preferable for de-
under noise. In fact, this label appears in all instances of scr tecting paradigmatic relationships between words (those that
scores as a top label, covering 26.9% of words in the clus- hold between words that do not co-occur themselves but oc-
ter. However, the rest of the top labels are either of a general cur in similar contexts, e.g. dog and cat) and at the same time
kind (entity, change) or not related to sound or movement there is evidence (Peirsman, Heylen, & Geeraerts, 2008) that
(handwriting, wound) and therefore we cannot consider the larger context is beneficial for picking out syntagmatic rela-
labeling result to be very strong. In other cases, the words are tionships that hold between words that often occur together
primarily subsumed by labels like entity and abstraction. (e.g. “crystal clear”). To our knowledge, the kinds of rela-
As explained in the Methods, we considered the possibility tionships that hold between phonesthemic words (in general
that clusters might be composed of several groups of words or depending on a given phonestheme) have not been system-
that do not all share the same semantic content. This is espe- atically investigated using such distinctions and further work
cially likely for numerous clusters (e.g. gr cluster even with on the influence of the kind of context useful for detecting
proper nouns removed contained 158 words). To counteract phonesthemic relatedness, in conjunction with experimental
this problem we examined how prior sub-clustering affects
the labeling results. The EM algorithm we used detected the 8 For the other phonesthemes, no sub-clusters were detected.
 work on similarity, could offer clues on this issue.
Our labeling results are somewhat disappointing but, given
the novelty of our approach, still highly informative. The fact
that we obtain better results for sn and str phonesthemes with
algorithms that favor specificity over coverage and that label-
ing is not fully successful with the remaining phonesthemes
are puzzling given the high support that we obtain for these
phonesthemes in our first experiment. We believe that there
are two possibilities that can explain this.
The first possibility is that our WordNet-based methodol-
ogy is not fully suited to discover the common semantic con-
tent that is present. WordNet does not allow for integrating
hypernymy tree chains across different parts of speech, which
might be vital for phonesthemes, a large proportion of which
are verbs. In addition, it does not make all the distinctions that
would be useful for phonesthemic studies, e.g. both scr and gr
are associated with kinds of sound but one is “unpleasant” and
the other one “threatening” – a distinction which is not part
of the WordNet taxonomy. On a more general note, hyper-
nymy might not be the most appropriate relation for all phon-
esthemes, e.g. snout and sneezing are not similar because they
are both a type of nose. Therefore, perhaps better labeling re-
sults could be achieved using a different semantic network,
such as ConceptNet9 , which allows for exploiting other than
merely “is a” relations.
The second possibility, which we cannot reject, is that there
is in fact no semantic core that unites phonestheme clusters
and that the statistical support obtained by Otis and Sagi and
in our first experiment is a result of a particular methodology.
This interpretation is suggested by the fact that in our second
experiment the overall coverage of phonesthemic words by
the semantic labels is relatively low. Qualitative examination
of the clusters also seems to show that they contain a lot of
variability. In the future, we plan to design stricter tests –
for example, comparing phonestheme clusters to clusters that
share a particular (non-phonesthemic) sub-string rather than
simply to a group of random words.
Ultimately, the aim of automatically detecting phones-
themes and their semantic content in a more objective, falsifi-
able way is, on the one hand, to help researchers interested in
iconicity to validate the existence of phonesthemes previously
reported in the literature, to possibly discover new phones-
themes, and to settle disputes over their particular meaning;
and on the other hand, to open the door to investigating fur-
ther the cognitive nature of the semantic relationships that
unite phonestheme clusters. This study constitutes a step in
this research programme.
References
Bellugi, U., & Klima, E. (1976). Two faces of sign: Iconic
and abstract. Annals of the New York Academy of Sciences,
280, 514-538.
Bergen, B. (2004). The psychological reality of phonaes-
themes. Language, 80(2), 291-311.
9 Freely available from http://conceptnet5.media.mit.edu.
View publication stats
Bird, S., Klein, E., & Loper, E. (2009). Natural language
processing with python: Analyzing text with the natural
language toolkit. O’Reilly Media.
Blust, R. A. (2003). The phonestheme n-in austronesian lan-
guages. Oceanic Linguistics, 42(1), 187-212.
Burnard, L. (Ed.). (2007). Reference Guide for the British
National Corpus (XML Edition). Oxford Universtity Com-
puting Services.
Drellishak, S. (2006). Statistical techniques for detecting
and validating phonesthemes. Unpublished master’s thesis,
University of Washington.
Fellbaum, C. (2005). Wordnet and wordnets. In Encyclopedia
of language and linguistics (Second ed.). Oxford: Elsevier.
Firth, J. (1930). Speech. London: Oxford University Press.
Hinton, L., Nichols, J., & Ohala, J. J. (2006a). Introduction.
In Sound symbolism. Cambridge University Press.
Hinton, L., Nichols, J., & Ohala, J. J. (Eds.). (2006b). Sound
symbolism. Cambridge University Press.
Hutchins, S. S. (1998). The psychological reality, variabil-
ity, and compositionality of english phonesthemes. Unpub-
lished doctoral dissertation, Emory University.
Marchand, H. (1969). The categories and types of present-
day english word-formation. Munich: C. H. Beck.
McDonald, S., & Ramscar, M. (2001). Testing the distribu-
tional hypothesis: The influence of context on judgements
of semantic similarity. In Proc. 23rd CogSci.
Ohala, J. J. (1994). The frequency codes underlies the sound
symbolic use of voice pitch. In Sound symbolism. CUP.
Otis, K., & Sagi, E. (2008). Phonaesthemes: A corpus-based
analysis. In Proc. 30th CogSci.
Peirsman, Y., Heylen, K., & Geeraerts, D. (2008). Size mat-
ters: Tight and loose context definitions in English word
space models. In Proc. ESSLLI Workshop on Distributional
Lexical Semantics.
Perniss, P., Thompson, R., & Vigliocco, G. (2010). Iconicity
as a general property of language: Evidence from spoken
and signed languages. Frontiers in Psychology, 1, 1-15.
Reid, R. (1967). Sound symbolism. T&A Constable.
Rhodes, R. (2006). Aural images. In Sound symbolism. CUP.
Sahlgren, M. (2005). An introduction to Random Indexing.
In Proc. Methods & Applications of Semantic Indexing.
Sahlgren, M. (2008). The distributional hypothesis. Italian
Journal of Linguistics, 20(1), 33-54.
Schütze, J. (1997). Ambiguity resolution in language learn-
ing. Stanford: CSLI Publications.
Shrum, L., & Lowrey, T. (2007). Sounds convey meaning:
The implications of phonetic symbolism for brand name
construction. In Psycholinguistic phenomena in marketing
communications. Mahwah: Erlbaum.
Widdows, D. (2003). Unsupervised methods for developing
taxonomies by combining syntactic and statistical informa-
tion. In Proc. HLT-NAACL.
Widdows, D., & Cohen, T. (2010). The semantic vectors
package: New algorithms and public tools for distributional
semantics. In Proc. 4th Int’l Conf. on Semantic Computing.